Gated Convolutional Recurrent Networks with Efficient Channel attention for Monaural Speech Enhancement
B. Manaswini, Venkata Adi Lakshmi, Loukya Chintha, N.Siva Kotaiah · 2023
Deep convolutional neural networks performance has been improved using the channel attention technique. Despite the fact that performance has improved, the model network’s complexity has grown. An efficient channel Attention layer, which only requires a small number of parameters and yields a definite performance boost, resolves the issue of the complexity and performance trade-off. By contrasting them with those of noisy speech, complex spectral mapping calculates the clear speech’s real and imaginary spectrograms. The simultaneous enhancement of phase and magnitude responses to speech. In order to improve monaural speech quality, the authors suggested a gated convolutional recurrent network with efficient channel attention (GCRN-ECA) for complex spectrum mapping. This network is based on multi- task learning. In the encoder and decoder, each layer is made out of dense blocks. Without dimensionality reduction, cross- channel interaction may be accomplished using the ECA module. While maintaining efficiency, a suitable cross-channel interaction can drastically reduce model complexity. Proposed findings show that the proposed GCRN-ECA fares better with the time of quality and understandability than the current convolutional neural networks (CNN). The STOI and PESQ of the suggested method are also significantly greater than those of spectral mapping and complicated ratio masking, to put it another way.